02Observe
See every run, every call, every cost.
Worksi records each model call, tool call, step and approval as a span and stitches them into runs, so you can follow an agent from the first prompt to the final outcome, live or weeks later.
- workflow_runsupport-ticket-resolutionone run, one trace
one run, one trace
- agent_steptriagetriage-agent
triage-agent
- llm_callgpt-5-minichat.completions · tokens, cost, latency
chat.completions · tokens, cost, latency
- agent_stepresearchresolver-agent
resolver-agent
- llm_callembeddingstext-embedding-3-small · tokens, cost
text-embedding-3-small · tokens, cost
- tool_callsearch_kbkb-vector-store
kb-vector-store
- tool_calllookup_orderorders-db · args hash
orders-db · args hash
- agent_steprefundrefund-agent
refund-agent
- llm_callgpt-5-nanochat.completions · refund eligibility
chat.completions · refund eligibility
- tool_callissue_refundpayments-api
payments-api
- agent_stepdraft_replyresolver-agent
resolver-agent
- llm_callgpt-5chat.completions · reply draft
chat.completions · reply draft
- agent_stepnotifyresolver-agent
resolver-agent
- tool_callsend_emailthe reply to the customer
the reply to the customer
One schema for every source.
Gateway calls, SDK spans, MCP tool calls and OpenTelemetry traces land in one span model, so dashboards, detection and the graph work the same whichever way you connect.
- Gateway
- An
llm_callfor every chat, Responses and embeddings call, with model, tokens, cost and latency. - SDKs
workflow_run,agent_step,tool_call,retrievalandhuman_actionaround your code. Model calls become children of the step.- MCP
- A
tool_callfor every tool an agent calls through the MCP gateway, with the server, an argument hash, the status and the latency. - OpenTelemetry
- GenAI spans mapped by
gen_ai.operation.name: model calls, tool calls and agent steps, in the same traces.
One run, one screen.
Open any trace for its span waterfall with timings, the failure location marked, and an inspector with model, tokens, cost, status, errors and the previews your capture mode allows.
1. Linked incident. The trace links straight to INC‑5.
2. The first failure, named at the top of the trace.
3. Run totals: duration, spans, tokens and cost.
Live, as it happens.
A streaming feed of model and tool calls with latency, tokens and cost, next to live counters for requests per minute, error rate, active runs and spend.
1. Every model and tool call as it happens.
2. Attributed to agent and workflow.
3. Status on every call.
Choose how much content is stored.
Every environment has a capture mode. A client can make it stricter per request, never looser.
| Mode | What is stored |
|---|---|
full | Input and output previews up to 2,000 characters. Payloads over 16 KiB go to object storage for 90 days. |
redacted (default) | The same previews, with sensitive values replaced by [REDACTED:TYPE]. |
metadata_only | No previews: hashes, counts, tokens, cost, timings, status and error types. |
What depends on what, and what breaks.
The AI Graph combines what you declared with what the traffic shows: workflows, agents, models, providers, tools and data sources, with call volumes and error rates on every edge.
- Impact mode highlights everything upstream and downstream of a node.
- Affected workflows, runs and cost at a glance.
- Agents, models, workflows and tools seen in traffic but never registered appear as discovered, ready to register or ignore.
- Rebuilt every five minutes from the last seven days of traffic.
- Included in Pro, Business and Enterprise, and in the trial.
Name the agent on every call.
Two optional headers attribute each request to an agent and a workflow, and your traces group themselves.
curl https://gateway.worksi.ai/v1/chat/completions \ -H "Authorization: Bearer $AIW_API_KEY" \ -H "X-AIW-Agent: support-bot" \ -H "X-AIW-Workflow: customer-chat" \ -H "Content-Type: application/json" \ -d '{"model": "gpt-5-mini", "messages": [{"role": "user", "content": "Hi"}]}'